Detect orientation of screw image

The screw as shown in the image is not completely horizontal. I would like to detect the orientation of screw and form a new image that can make sure that the screw is completely horizontal. Is there a way to do that?

[KSSV]

Read the pixels...get the coordinates...use polyfit to fit a straight line, get the slope and try to rotate the coordinates to make slope zero.

[Ang Xian Jia]

This is the pixels I read from top part of the screw. May I know what you mean by " try to rotate the coordinates to make slope zero" ? Thanks

Answer 1

[Stephan]

Using this script lets you know which angle to rotate. Then use this information to rotate the image and check the orientation again:

%% read image and find points that should lie on a straight line
img = rgb2gray(imread('20.bmp'));
res = detectHarrisFeatures(img,'MinQuality', 0.1, 'ROI', [1 386 762 20]);

% Extract coordinates of the points
x_vals = double(res.Location(:,1));
y_vals = double(res.Location(:,2));

% find slope and offset of the line formed by the points and use arctan to
% estimate the angle the srew is different from horizontal orientation
slope_and_offset = polyfit(x_vals, y_vals,1);
angle = atand(slope_and_offset(1));
fprintf('\nThe screw is oriented %.5f degree to the horizontal.\n', angle) 

%% rotate the image
img_rot = imrotate(img,angle);

% search points again on rotated image
res_new = detectHarrisFeatures(img_rot,'MinQuality', 0.1, 'ROI', [1 395 762 10]);

% Extract coordinates of the points
x_vals_rot = double(res_new.Location(:,1));
y_vals_rot = double(res_new.Location(:,2));

% find slope and offset of the line formed by the points and use arctan to
% estimate the angle the srew is different from horizontal orientation
slope_and_offset_rot = polyfit(x_vals_rot, y_vals_rot,1);
angle_rot = atand(slope_and_offset_rot(1));
fprintf('\nThe screw is oriented %.5f degree to the horizontal after rotation.\n\n', angle_rot) 

%% plot results
subplot(1,2,1)
hold on
title('Original Image')
imshow(img)
scatter(res.Location(:,1),res.Location(:,2),'or')
hold off

subplot(1,2,2)
hold on
title('Corrected Image')
imshow(img_rot)
scatter(res_new.Location(:,1),res_new.Location(:,2),'ob')
hold off

The keypoint is to define the ROI to the algorithm, so that you only detect points that are valid to reach high presicion.

[Ang Xian Jia]

For this method, it focuses on bottom screw thread. What if the angle of polyfit of top thread with horizontal axis is different with that of bottom thread? Is it possible to use the centre axis of screw to rotate the image to become horizontal?

[Stephan]

If we apply the same operation on both sides as i did for the lower part, it should be able to estimate the center line, even if the screw would be conic.

[Ang Xian Jia]

Hi Sir Stephan, u means that I should take the average of two polyfit line (top and bottom) to form the centerline?

[Stephan]

Yes, this is what i would try to do.

Answer 2

[Image Analyst]

An alternative way is to use the radon transform. Basically the radon transform makes projections along a list of angles that you specify. Then you can look at the projections image to identify which angle (angles are columns in the image) is the best one (it will be the column where the brightest value is).

See attached demo where I rotate the football demo image.

[Ang Xian Jia]

Hi Image Analyst, why the code would not work if I use the image below for the code? Do I need to make any modification? Why we are extracting only the red channel? The reason I'm using different kind of image is because I'm curious whether the code will always orientate different image to horizional position or not. I'm still trying to understand the codes given. Thanks.

[Image Analyst]

I extracted the red channel because radon() works on a grayscale image, not a color RGB image. You might either call rgb2gray() or use the green channel (since cameras have twice as many green pixels as red or blue pixels).

You have an image capture problem. Bad illumination, bad background, bad lens, bad geometry.

You need to

1. Use a uniform white lightbox to create a silohette image.
2. Use a telecentric lens so magnification does not change with distance from lens.
3. Aim camera normal to screw axis, not at an angle like you have.

[Ang Xian Jia]

Yeah I understand that. My uni has only limited and old equipment for machine vision course yet I must have a clear screw image. I will still try to recapture the image. Thanks for the suggestion.

[Ang Xian Jia]

May I know why peak of radon transform image is the parameter used to rotate the image to ensure the perfectly horizontal and vertical orientation? I have tried to read the documentation but still dont understand.

[Image Analyst]

The radon transform is the sum of the image along a certain direction. Kind of like an x-ray. So for the football, the highest value will be when the direction of projection lies right along the longest axis of the football, which is exactly the angle we want to use for rotation to align with the main image axes.

[Ang Xian Jia]

Hi image Analyst,

May I know why the screw will not orientated perfect horizontally for this kind of image?

[Image Analyst]

You should be looking for the minimum sum since the object is dark on a bright background. Plus you're going to have to add an offset because the longest dark stretch is along the diagonal, not along the axis of the bolt.

An alternate way is to just take all the points with find() and then fit them to a line with polyfit. Then get the angle from the arctangent of the slope and then call imrotate(). Something like (untested):

[y, x] = find(binaryImage);
coefficients = polyfit(x, y, 1);
angle = arctand(coefficients(1));
rotatedImage = imrotate(grayImage, -angle);
imshow(rotatedImage);

I've used an approach like this for rectangular objects in the past, and it worked well.

[Ang Xian Jia]

Understood. Thank you very much, Image Analyst.